Published by the IEEE Computer Society H I S T O R I E S A N D F U T U R E S Editors: Robert R. Hoffman, Jeffrey M. Bradshaw, and Kenneth M. Ford, Institute for Human and Machine Cognition, rhoffman@ihmc.us H U M A N - C E N T E R E D C O M P U T I N G that the robot had complex perception and rea- soning skills equivalent to a child and that robots were subservient to humans. Although the laws were simple and few, the stories attempted to dem- onstrate just how diffcult they were to apply in various real-world situations. In most situations, although the robots usually behaved logically, they often failed to do the right thing, typically because the particular context of application re- quired subtle adjustments of judgment on the part of the robot (for example, determining which law took priority in a given situation, or what consti- tuted helpful or harmful behavior). The three laws have been so successfully incul- cated into the public consciousness through enter- tainment that they now appear to shape societys expectations about how robots should act around humans. For instance, the media frequently refer to humanrobot interaction in terms of the three laws. Theyve been the subject of serious blogs, events, and even scientifc publications. The Sin- gularity Institute organized an event and Web site, Three Laws Unsafe, to try to counter pub- lic expectations of robots in the wake of the movie I, Robot. Both the philosophy 1 and AI 2 commu- nities have discussed ethical considerations of ro- bots in society using the three laws as a reference, with a recent discussion in IEEE Intelligent Sys- tems. 3 Even medical doctors have considered ro- botic surgery in the context of the three laws. 4 With few notable exceptions, 5,6 there has been relatively little discussion of whether robots, now or in the near future, will have suffcient percep- tual and reasoning capabilities to actually follow the laws. And there appears to be even less serious discussion as to whether the laws are actually vi- able as a framework for humanrobot interaction, outside of cultural expectations. Following the defnitions in Moral Machines: Teaching Robots Right from Wrong, 7 Asimovs laws are based on functional morality, which as- sumes that robots have suffcient agency and cog- nition to make moral decisions. Unlike many of his successors, Asimov is less concerned with the details of robot design than in exploiting a clever literary device that lets him take advantage of the large gaps between aspiration and reality in robot autonomy. He uses the situations as a foil to ex- plore issues such as the ambiguity and cultural dependence of lan- guage and behaviorfor example, whether what appears to be cruel in the short run can actually become a kindness in the longer term; social utilityfor instance, how different indi- viduals roles, capabilities, or backgrounds are valuable in different ways with respect to each other and to society; and the limits of technologyfor example, the im- possibility of assuring timely, correct actions in all situations and the omnipresence of trade-offs. In short, in a variety of ways the stories test the lack of resilience in humanrobot interactions. The assumption of functional morality, while ef- S ince their codifcation in 1947 in the col- lection of short stories I, Robot, Isaac Asi- movs three laws of robotics have been a staple of science fction. Most of the stories assumed Robin R. Murphy, Texas A&M University David D. Woods, Ohio State University Beyond Asimov: The Three Laws of Responsible Robotics JULY/AUGUST 2009 www.computer.org/intelligent 15 fective for entertaining storytelling, neglects operational morality. Oper- ational morality links robot actions and inactions to the decisions, as- sumptions, analyses, and investments of those who invent and make ro- botic systems and of those who com- mission, deploy, and handle robots in operational contexts. No matter how far the autonomy of robots ultimately advances, the important challenges of these accountability and liability link- ages will remain. 8 This essay reviews the three laws and briefy summarizes some of the practical shortcomingsand even dangersof each law for framing humanrobot relationships, includ- ing reminders about what robots cant do. We then propose an alter- native, parallel set of laws based on what humans and robots can real- istically accomplish in the foresee- able future as joint cognitive systems, and their mutual accountability for their actions from the perspectives of human-centered design and human robot interaction. Applying Asimovs Laws to Todays Robots When we try to apply Asimovs laws to todays robots, we immediately run into problems. Just as for Asi- mov in his short stories, these prob- lems arise from the complexities of situations where we would use ro- bots, the limits of physical systems acting with limited resources in un- certain changing situations, and the interplay between the different social roles as different agents pursue mul- tiple goals. First Law Asimovs frst law of robotics states, A robot may not injure a human be- ing or, through inaction, allow a hu- man being to come to harm. This law is already an anachronism given the militarys weaponization of ro- bots, and discussions are now shifting to the question of whether weaponized robots can be humane. 9,10 Such weaponization is no longer limited to situations in which humans remain in the loop for control. The South Ko- rean government has published vid- eos on YouTube of robotic border- security guards. Scenarios have been proposed where it would be permis- sible for a military robot to fre upon anything moving (presumably target- ing humans) without direct human permission. 11 Even if current events hadnt made the law irrelevant, its moot because robots cannot infallibly recognize hu- mans, perceive their intent, or reli- ably interpret contextualized scenes. A quick review of the computer vi- sion literature shows that scientists continue to struggle with many fun- damental perceptual processes. Cur- rent commercial security packages for recognizing the face of a person standing in a fxed position continue to fall short of expectations in prac- tice. Many robots that recognize humans use indirect cues such as heat and motion, which only work in constrained contexts. These prob- lems confrm Norbert Wieners warn- ings about such failure possibilities. 8
Just as he envisioned many years ago, todays robots are literal-minded agentsthat is, they cant tell if their world model is the world theyre really in. All this aside, the biggest problem with the frst law is that it views safety only in terms of the robotthat is, the robot is the responsible safety agent in all matters of humanrobot interaction. While some speculate on what it would mean for a robot to be able to discharge this responsibility, there are serious practical, theoreti- cal, social-cognitive, and legal limi- tations. 8,12 For example, from a legal perspective the robot is a product, so its not the responsible agent. Rather, the robots owner or manufacturer is liable for its actions. Unless robots are granted a person-equivalent sta- tus, somewhat like corporations are now legally recognized as individual entities, its diffcult to imagine stan- dard product liability law not apply- ing to them. When a failure occurs, violating Asimovs frst law, the hu- man stakeholders affected by that failure will engage in the processes of causal attribution. Afterwards, theyll see the robot as a device and will look for the person or group who set up or instructed the device erro- neously or who failed to supervise (that is, stop) the robot before harm occurred. Its still commonplace af- ter accidents for manufacturers and organizations to claim the result was due only to human error, even when the system in question was operating autonomously. 8,13 Accountability is bound up with the way we maintain our social re- lationships. Human decision-making always occurs in a context of expec- tations that one might be called to account for his or her decisions. Ex- pectations for whats considered an adequate explanation and the con- sequences for people when their explanation is judged inadequate are Asimovs laws are based on functional morality, which assumes that robots have suffcient agency and cognition to make moral decisions. 16 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS critical parts of accountability sys- temsa reciprocating cycle of being prepared to provide an accounting for ones actions and being called by oth- ers to provide an account. To be con- sidered moral agents, robots would have to be capable of participating personally in this reciprocating cycle of accountabilityan issue that, of course, concerns more than any sin- gle agents capabilities in isolation. Second Law Asimovs second law of robotics states, A robot must obey orders given to it by human beings, except where such orders would confict with the frst law. Although the law itself takes no stand on how humans would give or- ders, Asimovs robots relied on their understanding of verbal directives. Unfortunately, robust natural lan- guage understanding still continues to lie just beyond the frontiers of todays AI. 14 Its true that, after decades of re- search, computers can now construct words from phonemes with some con- sistencyas improvements in voice dictation and call centers attest. Lan- guage-understanding capabilities also work well for specifc types of well- structured tasks. However, the goal of meaningful machine participation in open-ended conversational con- texts remains elusive. Additionally, we must account for the fact that not all directives are given verbally. Humans use gestures and add affect through body posture, facial expressions, and motions for clarifcation and empha- sis. Indeed, high-performance, expe- rienced teams use highly pointed and coded forms of verbal and nonverbal communication in fuid, interdepen- dent, and idiosyncratic ways. Whats more interesting about the second law from a humanrobot in- teraction standpoint is that at its core, it almost captures the more important idea that intelligent robots should no- tice and take stock of humans (and that the people robots encounter or interact with can notice pertinent as- pects of robots behavior). 15 For ex- ample, is it acceptable for a robot to merely not hit a person in a hospi- tal hall, or should it conform to so- cial convention and acknowledge the person in some way (excuse me or a nod of a camera pan-tilt)? Or if a robot operating in public places in- cluded two-way communication de- vices, could a bystander recognize that the robot provided a means to report a crime or a fre? Third Law The third law states, A robot must protect its own existence as long as such protection does not confict with the frst or second law. Because to- days robots are expensive, youd think designers would be naturally motivated to incorporate some form of the third law into their products. For example, even the inexpensive iRobot Roomba detects stairs that could cause a fatal fall. Surprisingly, however, many expensive commercial robots lack the means to fully protect their owners investment. An extreme example of this is in the design of robots for military ap- plications or bomb squads. Such ro- bots are designed to be teleoperated by a person who bears full responsi- bility for all safety matters. Human- factors studies show that remote operators are immediately at a dis- advantage, working through a me- diated interface with a time delay. Worse yet, remote operators are re- quired to operate the robot through poor humancomputer interfaces and in contexts where the operator can be fatigued, overloaded, or under high stress. As a result, when an abnormal event occurs, they may be distracted or not fully engaged and thus might not respond adequately in time. The result for a robot is akin to expecting an astronaut on a planets surface to request and wait for permission from mission control to perform even sim- ple refexes such as ducking. What is puzzling about todays lim- ited attempts to conform to the third law is that there are well-established technological solutions for basic ro- bot survival activities that work for autonomous and human-controlled robots. For instance, since the 1960s weve had technology to assure guarded motion, where the human drives the robot but onboard software will not allow the robot to make po- tentially dangerous moves (for exam- ple, collide with obstacles or exceed speed limits or boundaries) without explicit orders (an implicit invocation of the second law). By the late 1980s, guarded motion was encapsulated into tactical reactive behaviors, essen- tially giving robots refexes and tac- tical authority. Perhaps the most im- portant reason that guarded motion and refexive behaviors havent been more widely deployed is that they re- quire additional sensors, which would add to the cost. This increase in cost may not appear to be justifed to cus- tomers, who tend to be wildly over- confdent that trouble and complexi- ties outside the bounds of expected behavior rarely arise. The goal of meaningful machine participation in open-ended conversational contexts remains elusive. JULY/AUGUST 2009 www.computer.org/intelligent 17 The Alternative Three Laws of Responsible Robotics To address the diffculties of apply- ing Asimovs three laws to the cur- rent generation of robots while re- specting the laws general intent, we suggest the three laws of responsible robotics. Alternative First Law Our alternative to Asimovs frst law is A human may not deploy a ro- bot without the humanrobot work system meeting the highest legal and professional standards of safety and ethics. Since robots are indeed sub- ject to safety regulations and liability laws, the requirement of meeting legal standards for safety would seem self- evident. For instance, the medical- device community has done extensive research to validate robot sensing of scalpel pressures and tissue contact parameters, and it invests in failure mode and effect analyses (consistent with FDA medical-device standards). In contrast, mobile roboticists have a somewhat infamous history of disregarding regulations. For ex- ample, robot cars operating on pub- lic roads, such as those used in the DARPA Urban Grand Challenge, are considered by US federal and state transportation regulations as experi- mental vehicles. Deploying such vehi- cles requires voluminous and tedious permission applications. Regrettably, the 1995 CMU No Hands Across America team neglected to get all appropriate permissions while driv- ing autonomously from Pittsburgh to Los Angeles, and were stopped in Kansas. The US Federal Aviation Ad- ministration makes a clear distinc- tion between fying unmanned aerial vehicles as a hobby and fying them for R&D or commercial practices, effectively slowing or stopping many R&D efforts. In response to these diffculties, a culture preferring for- giveness to permission has grown up in some research groups. Such at- titudes indicate a poor safety culture at universities that could, in turn, propagate to government or industry. On the positive side, the robot com- petitions sponsored by the Associa- tion for Unmanned Vehicle Systems International are noteworthy in their insistence on having safe areas of op- eration, clear emergency plans, and safety offcers present. Meeting the minimal legal require- ments is not enoughthe alternative frst law demands the highest profes- sional ethics in robot deployment. A failure or accident involving a robot can effectively end an entire branch of robotics research, even if the op- erators arent legally culpable. Re- sponsible communities should proac- tively consider safety in the broadest sense, and funding agencies should fnd ways to increase the priority and scope of research funding specifcally aimed at relevant legal concerns. The highest professional ethics should also be applied in product de- velopment and testing. Autonomous robots have known vulnerabilities to problems stemming from interrupted wireless communications. Signal re- ception is impossible to predict, yet robust return to home if signal lost and stop movement if GPS lost functionality hasnt yet become an expected component of built-in robot behavior. This means robots are oper- ating counter to reasonable and pru- dent assumptions. Worse yet, when theyre operating experimentally, ro- bots often encounter unanticipated factors that affect their control. Sim- ply saying an unfortunate event was unpredictable doesnt relieve the de- signers of responsibility. Even if a specifc disturbance is unpredictable in detail, the fact that there will be disturbances is virtually guaranteed, and designing for resilience in the face of these is fundamental. As a matter of professional com- mon sense, robot design should start with safety frst, then add the inter- esting software and hardware. Ro- bots should carry black boxes or recorders to show what they were doing when a disturbance occurred, not only for the sake of an accident investigation but also to trace the ro- bots behavior in context to aid diag- nosis and debugging. There should be a formal safety plan and checklists for contingencies. These do not have to be extensive and time consuming to be effective. A litmus test for de- velopers might be If a group of ex- perts from the IEEE were to write about your robot after an accident, what would they say about system safety and your professionalism? Fundamentally, the alternative frst law places responsibility for safety and effcacy on humans within the larger social and environmental con- text in which robots are developed, deployed, and operated. Alternative Second Law As an alternative to Asimovs sec- ond law, we propose the follow- ing: A robot must respond to hu- mans as appropriate for their roles. The capability to respond appropri- atelyresponsivenessmay be more The fact that there will be disturbances is virtually guaranteed, and designing for resilience in the face of these is fundamental. 18 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS important to humanrobot interac- tion than the capability of autonomy. Not all robots will be fully autono- mous over all conditions. For exam- ple, a robot might be constrained to follow waypoints but will be expected to generate appropriate responses to people it encounters along the way. Responsiveness depends on the social environment, the kinds of people and their expectations that a robot might encounter in its work envelope. Rather than assume the relationship is hier- archical with the human as the supe- rior and the robot as the subordinate so that all communication is a type of order, the alternative second law states that robots must be built so that the interaction fts the relationships and roles of each member in a given environment. The relationship deter- mines the degree to which a robot is obligated to respond. It might ignore a hacker completely. Orders exceeding the authority of the speaker might be disposed of politely (please have your superior confrm your request) or with a warning (interference with a law enforcement robot may be a viola- tion). Note that defning appropri- ate response may address concerns about robots being abused. 16 The relationship also determines the mode of the response. How the robot signals or expresses itself should be consistent with that relationship. Ca- sual relationships might rely on natu- ral language, whereas trained teams performing specifc tasks could coor- dinate activities through other signals such as body position and gestures. The requirement for responsive- ness captures a new form of autonomy (not as isolated action but the more diffcult behavior of engaging appro- priately with others). However, devel- oping robots capability for respon- siveness requires a signifcant research effort, particularly in how robots can perceive and identify the different members, roles, and cues of a social environment. Alternative Third Law Our third law is A robot must be endowed with suffcient situated au- tonomy to protect its own existence as long as such protection provides smooth transfer of control to other agents consistent with the frst and second laws. This law specifes that a humanrobot system should be able to transition smoothly from whatever degree of autonomy or roles the ro- bots and humans were inhabiting to a new control relationship given the na- ture of the disruption, impasse, or op- portunity encountered or anticipated. When developers focus narrowly on the goal of isolated autonomy and fall prey to overconfdence by underesti- mating the potential for surprises to occur, they tend to minimize the im- portance of transfer of control. But bumpy transfers of control have been noted as a basic diffculty in human interaction with automation that can contribute to failures. 17 The alternative third law addresses situated autonomy and smooth trans- fer of control, both of which interact with the prescriptions of the other laws. To be consistent with the second law requires that humans in a given role might not always have complete control of the robot (for example, when conditions require very short reaction times, a pilot may not be al- lowed to override some commands generated by algorithms that attempt to provide envelope protection for the aircraft). This in turn implies that an aspect of the design of roles is the identifcation of classes of situations that demand transfer of control, so that the exchange processes can be specifed as part of roles. This is when the human takes control from the ro- bot for a specialized aspect of the mis- sion in anticipation of conditions that will challenge the limits of the robots capabilities, or in an emergency. De- cades of human factors research on human out-of-the-loop control prob- lems, handling of anomalies, cascades of disturbances, situation awareness, and autopilot/pilot transfer of control can inform such designs. To be consistent with the frst law requires designers to explicitly ad- dress what is the appropriate situated autonomy (for example, identifying when the robot is better informed or more capable than the human owing to latency, sensing, and so on) and to provide mechanisms that permit smooth transfer of control. To disre- gard the large body of literature on resilience and failure due to bumpy transfer of control would violate the designers ethical obligation. The alternative second and third laws encourage some forms of in- creased autonomy related to respon- siveness and the ability to engage in various forms of smooth transfer of control. To be able to engage in these activities with people in vari- ous roles, the robot will need more situated intelligence. The result is an irony that has been noted before: in- creased capability for autonomy and authority leads to the need to partici- pate in more sophisticated forms of coordinated activity. 8 Increased capability for autonomy and authority leads to the need to participate in more sophisticated forms of coordinated activity. JULY/AUGUST 2009 www.computer.org/intelligent 19 Discussion Our critique reveals that robots need two key capabilities: responsiveness and smooth transfer of control. Our proposed alternative laws remind ro- botics researchers and developers of their legal and professional responsi- bilities. They suggest how people can conduct humanrobot interaction re- search safely, and they identify criti- cal research questions. Table 1 places Asimovs three laws side by side with our three alternative laws. Asimovs laws assume functional moralitythat robots are capable of making (or are permitted to make) their own decisionsand ignore the legal and professional responsibility of those who design and deploy them (operational morality). More impor- tantly for humanrobot interaction, Asimovs laws ignore the complexity and dynamics of relationships and re- sponsibilities between robots and peo- ple and how those relationships are expressed. In contrast, the alternative three laws emphasize responsibility and resilience, starting with enlight- ened, safety-oriented designs (alter- native frst law), then adding respon- siveness (alternative second law) and smooth transfer of control (alternative third law). The alternative laws are designed to be more feasible to implement than Asimovs laws given current technol- ogy, although they also raise critical questions for research. For example, the alternative frst law isnt concerned with technology per se but with the need for robot developers to be aware of human systems design principles and to take responsibility proac- tively for the consequences of errors and failures in humanrobot systems. Standard tools from the aerospace, medical, and chemical manufacturing safety cultures, including training, formal processes, checklists, black boxes, and safety offcers, can be ad- opted. Network and physical security should be incorporated into robots, even during development. The alternative second and third laws require new research directions for robotics to leverage and build on existing results in social cognition, cognitive engineering, and resilience engineering. The laws suggest that the ability for robots to express re- lationships and obligations through social roles will be essential to all humanrobot interaction. For exam- ple, work on entertainment robots and social robots provides insights about how robots can express emo- tions or affect appropriate to people they encounter. The extensive litera- ture from cognitive engineering on transfer of control and general human out-of-the-loop control problems can be redirected at robotic systems. The techniques for resilience engineering are beginning to identify new control architectures for distributed, multi- echelon systems that include systems that include robots. The fundamental difference be- tween Asimovs laws, which focus on robots functional morality and full moral agency, and the alternative laws, which focus on system respon- sibility and resilience, illustrates why the robotics community should re- sist public pressure to frame current humanrobot interaction in terms of Asimovs laws. Asimovs laws dis- tract from capturing the diversity of robotic missions and initiative. Un- derstanding these diversities and complexities is critical for designing the right interaction scheme for a given domain. Ironically, Asimovs laws really are robot-centric because most of the ini- tiative for safety and effcacy lies in the robot as an autonomous agent. The alternative laws are human- centered because they take a systems approach. They emphasize that responsibility for the consequences of robots successes and failures lies in the human groups that have a stake in the robots activities, and capable robotic agents still exist in a web of dynamic social and cogni- tive relationships. Ironically, meeting the requirements of the alternative laws leads to the need for robots to be more capable agentsthat is, more responsive to others and better at interaction with others. We propose the alternative laws as a way to stimulate debate about ro- bots accountability when their ac- tions can harm people or human interests. We also hope that these Table 1. Asimovs laws of robotics versus the alternative laws of responsible robotics Asimovs laws Alternative laws 1 A robot may not injure a human being or, through inaction, allow a human being to come to harm. A human may not deploy a robot without the humanrobot work system meeting the highest legal and professional standards of safety and ethics. 2 A robot must obey orders given to it by human beings, except where such orders would conflict with the first law. A robot must respond to humans as appropriate for their roles. 3 A robot must protect its own existence as long as such protection does not conflict with the first or second law. A robot must be endowed with sufficient situated autonomy to protect its own existence as long as such protection provides smooth transfer of control to other agents consistent the first and second laws. 20 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS laws can serve to direct R&D to en- hance humanrobot systems. Finally, while perhaps not as entertaining as Asimovs laws, we hope the alterna- tive laws of responsible robotics can better communicate to the general public the complex mix of opportu- nities and challenges of robots in to- days world. Acknowledgments We thank Jeff Bradshaw, Cindy Bethel, Jenny Burke, Victoria Groom, and Leila Takayama for their helpful feedback and Sung Huh for additional references. The second authors contribution was based on participation in the Advanced Decision Ar- chitectures Collaborative Technology Alli- ance, sponsored by the US Army Research Laboratory under cooperative agreement DAAD19-01-2-0009.
References 1. S.L. Anderson, Asimovs Three Laws of Robotics and Machine Metaethics, AI and Society, vol. 22, no. 4, 2008, pp. 477493. 2. A. Sloman, Why Asimovs Three Laws of Robotics are Unethical, 27 July 2006; www.cs.bham.ac.uk/research/ projects/cogaff/misc/asimov-three-laws. html. 3. C. Allen, W. Wallach, and I. Smit, Why Machine Ethics? IEEE Intelligent Systems, vol. 21, no. 4, 2006, pp. 1217. 4. M. Moran, Three Laws of Robotics and Surgery, J. Endourology, vol. 22, no. 8, 2008, pp. 15571560. 5. R. Clarke, Asimovs Laws of Robotics: Implications for Information Technol- ogy Part 1, Computer, vol. 26, no. 12, 1993, pp. 5361. 6. R. Clarke, Asimovs Laws of Robotics: Implications for Information Technol- ogy Part 2, Computer, vol. 27, no. 1, 1994, pp. 5766. 7. W. Wallach and C. Allen, Moral Ma- chines: Teaching Robots Right from Wrong, Oxford Univ. Press, 2009. 8. D. Woods and E. Hollnagel, Joint Cognitive Systems: Patterns in Cogni- tive Systems Engineering, Taylor and Francis, 2006. 9. R.C. Arkin and L. Moshkina, Le- thality and Autonomous Robots: An Ethical Stance, Proc. IEEE Intl Symp. Technology and Society (ISTAS 07), IEEE Press, 2007, pp. 13. 10. N. Sharkey, The Ethical Frontiers of Robotics, Science, vol. 322, no. 5909, 2008, pp. 18001801. 11. M.F. Rose et al., Technology Develop- ment for Army Unmanned Ground Vehicles, Natl Academy Press, 2002. 12. D. Woods, Conficts between Learn- ing and Accountability in Patient Safety, DePaul Law Rev., vol. 54, 2005, pp. 485502. 13. S.W.A. Dekker, Just Culture: Bal- ancing Safety and Accountability, Ashgate, 2008. 14. J. Allen et al., Towards Conversa- tional HumanComputer Interaction, AI Magazine, vol. 22, no. 4, 2001, pp. 2738. 15. J.M. Bradshaw et al., Dimensions of Adjustable Autonomy and Mixed- Initiative Interaction, Agents and Computational Autonomy: Potential, Risks, and Solutions, M. Nickles, M. Rovatsos, and G. Weiss, eds., LNCS 2969, Springer, 2004, pp. 1739. 16. B. Whitby, Sometimes Its Hard to Be a Robot: A Call for Action on the Ethics of Abusing Artifcial Agents, Interacting with Computers, vol. 20, no. 3, 2008, pp. 326333. 17. D.D. Woods and N. Sarter, Learning from Automation Surprises and Going Sour Accidents, Cognitive Engineer- ing in the Aviation Domain, N.B. Sarter and R. Amalberti, eds., Natl Aeronau- tics and Space Administration, 1998. Robin Murphy is the Raytheon Profes- sor of Computer Science and Engineer- ing at Texas A&M University. Contact him at murphy@cse.tamu.edu. David D. Woods is a professor in the Human Systems Integration section of the Department of Integrated Systems Engineering at Ohio State University. Contact him at woods.2@osu.edu. IEEE Intelligent Systems delivers the latest peer-reviewed research on all aspects of artificial intelligence, focusing on practical, fielded applications. Contributors include leading experts in Intelligent Agents The Semantic Web Natural Language Processing Robotics Machine Learning I E E E Visit us on the Web at www.computer.org/intelligent THE #1 ARTIFICIAL INTELLIGENCE MAGAZINE!